Expressions

Expressions let characters “emote” visually during chats. After a character sends a message, NarratrixAI classifies the emotional tone and shows the best‑matching image from that character’s Expression Pack. Set up expressions per Character and drive them with a dedicated Chat Template that runs a lightweight classification prompt.

In the Expressions panel:

• Click Configure to edit prompts and model settings for expression classification.
• Click Generate to classify the most recent character message on demand.
• Toggle Auto to reclassify automatically when new messages arrive (runs on a short interval).

Tip
With Auto enabled, you can also select text in the transcript to classify only the selected content.

---

How It Works

• Each character has an Expression Pack (a set of images, one per expression name such as neutral, happy, angry).
• The Expressions panel sends a short prompt to your chosen model, asking it to return exactly one label from the character’s available expressions.
• NarratrixAI matches the model’s answer to the closest expression name and updates the image. If no close match is found, it falls back to neutral.

---

Setting Up Expressions

1. Choose or create a Chat Template dedicated to expressions.

   • Use a minimal or empty Prompt Format Template to avoid extra text.
   • Recommended Response Length: 60 (keeps answers to a single label).

2. Configure the System Prompt. It will replace any Prompt Format Template section.

   • Describe the job: classify tone/emotion based on character personality, recent message, and available labels.
   • Emphasize: “Return exactly one expression label from the provided list.”

3. Configure the User Prompt.

   • Provide the character name and personality.
   • Include the latest chat text to classify.
   • Include the list of allowed expressions.
   • Optionally include the last chosen expression to discourage repetition.

4. Optional: Prefill with a Custom Prompt.

   • Add a Custom Prompt with Role: Character, Position: Bottom of Conversation, Content: “The Expression is:” to hint single‑token completions (only helps on models that support prefilling/continuations).

5. Save Changes, then click Generate to test on the latest character message. Switch to Auto when satisfied.

---

Placeholders

When configuring Expression prompts, the following placeholders are available in addition to standard ones:

• {{expression.list}} — a comma‑separated list of the character’s available expression names.
• {{expression.last}} — the last expression chosen for this character.
• Common placeholders that are useful here:
  • {{char}}, {{character.name}}
  • {{character.personality}}
  • {{chat.message}} — the latest character message to classify

---

Example Prompt Snippets

System Prompt (excerpt):

• You classify character emotions and reply with exactly one label from the provided list.
• Stay consistent with the character’s personality and the conversation context.
• Do not explain your choice. Output only the label.

User Prompt (excerpt):

• Analyze: {{char}}
• Personality: {{character.personality}}
• Text to classify: {{chat.message}}
• Expression list: {{expression.list}}
• Last expression: {{expression.last}}

Inference suggestions:

• Temperature: 0.3–0.7 (lower favors consistent single‑label outputs)
• Keep penalties modest to avoid forcing novel labels

---

Expected Model Output

• Ideal: a single token or single word that exactly matches one item from {{expression.list}} (e.g., happy).
• Acceptable: a short phrase where the first word matches a label (NarratrixAI will fuzzy‑match).
• Fallback: if no match is found, NarratrixAI uses neutral.

Keep expression names short and distinct: neutral, happy, sad, angry, surprised, thinking.

---

Troubleshooting

• The image doesn’t change

  • Confirm the character has an Expression Pack and names match your prompt list.
  • Check that Auto is on (for continuous updates) or click Generate after a new message.

• The model returns sentences instead of a single label

  • Lower Response Length to ~60.
  • Tighten prompts to “Output only one label from the list—no explanation.”
  • Reduce Temperature.
  • Add a Prefill to adhere the model behavior.

• Wrong emotions get picked

  • Make expression names more distinct (avoid similar labels like happy vs cheerful).
  • Include {{expression.last}} and instruct the model to avoid repeats unless strongly indicated.

• Updated images aren’t appearing

  • If you changed files on disk, open the character, click Refresh in the Expression Pack, then try again.

---

Best Practices

• Use consistent naming across characters so one template works for all.
• Store a dedicated “Expressions” Chat Template separate from your main roleplay template. This avoids leaking long context into the classifier.